Small Language Models and Phi-3: A Guide 

Small language models (SLMs) are a powerful tool that has gained significant attention in recent years. They are used to generate text, images, and other forms of content with great precision. In this blog post, we will explore the concept of small language models and how you can use them using the Phi-3 cookbook.

What are Small Language Models?

Small language models are a type of neural network that is trained on a large corpus of text data. They are designed to generate new text that is similar to the training data, but they can also be used for tasks like image generation and more.

How do Small Language Models Work?

Small language models work by predicting the next word in a sequence based on the previous words. This process is known as autoregressive prediction. The model learns to generate text by repeatedly predicting the next word in a sequence and comparing its predictions with the actual word.

The Phi-3 Cookbook

The Phi-3 cookbook is an open-source repository that provides a collection of code examples for working with small language models. It includes tutorials, recipes, and other resources that will help you learn how to use these powerful models in your own projects.

What can you Learn from the Phi-3 Cookbook?

The Phi-3 cookbook is a great resource for technical students who want to learn more about small language models. Here are some of the things you can learn from it:

Generating Text with Small Language Models

You can use the Phi-3 cookbook to generate text using small language models. You can learn how to train your own model on a specific dataset, or you can use one of the pre-trained models that is included in the repository. Once you have generated some text, you can use it for tasks like summarization, machine translation, and more.

Generating Images with Small Language Models

You can also use small language models to generate images. The Phi-3 cookbook provides examples of how to use these models to create images from text descriptions. This is a great way to learn about computer vision and image generation.

Applying Small Language Models in Real-World Scenarios

The Phi-3 cookbook includes examples of how small language models can be applied in real-world scenarios. You can learn about how these models are used in applications like chatbots, virtual assistants, and more. This will give you a better understanding of the practical applications of this technology.

Conclusion

Small language models are a powerful tool that is becoming increasingly popular in the tech industry. The Phi-3 cookbook is an excellent resource for technical students who want to learn more about these models and how they can be used in their own projects. By working through the recipes and tutorials in this repository, you will gain valuable knowledge and skills that will help you succeed in your future career.

​Microsoft Tech Community – Latest Blogs –Read More 

Introduction

Azure Load Testing is a cloud-based service that lets you test the performance and scalability of your web applications and APIs. You can create and run load tests, analyze and compare the results, and define test fail criteria based on metrics like response time, error percentage, and throughput.

Response time is the time your application takes to respond to a user request. It depends on factors like network latency, server load, and application logic. To measure the performance of your application, you need to look at the response time distribution across different percentiles, which show how many requests are completed within a certain time limit. For example, the 90th percentile response time means that 90% of the requests are completed within that time, and the remaining 10% take longer.

What’s new in Azure Load Testing?

Until now, Azure Load Testing only supported three percentiles for response time analysis: Min, Max, Average, Median(50th), 90th, 95th, and 99th. These percentiles are common in the industry, but they may not match your SLA requirements. For instance, you may want to track the response time at the 75th percentile, which is a standard for web performance, or the 99.9th percentile, which is a stricter measure of the tail latency.

That’s why we are happy to announce that Azure Load Testing now supports more percentiles for response time analysis. You can configure the percentiles you want for response time metrics in Azure Load Testing and meet your SLA goals. You can choose from these percentiles:

Min
Max
Average
Median (50th)
75th
90th
95th
96th
97th
98th
99th
99.9th
99.99th

You can use these percentiles in different ways to analyze and compare your test results. For example, you can:

Define test fail criteria based on the percentiles you want for response time metrics.
Filter the ‘Sampler Statistics’ UI by the percentiles you want for response time metrics.
Filter the response time graphs by the percentiles you want for response time metrics.

How to configure percentiles in Azure Load Testing?

Configuring percentiles in Azure Load Testing is easy and intuitive. You can do it in two ways:

Using the Azure portal
Using the YAML file

Using the Azure portal, you can configure the percentiles for response time metrics in the test criteria and the test results pages. In the test criteria page, you can select the aggregate function for the response time metric from the drop-down list that includes all the percentiles as shown below.

In the test results page, you can filter the sampler statistics UI and the response time graphs by the percentiles you want using the filters.

Using the YAML file, you can configure the percentiles for response time when running tests from your CI/CD pipelines. You can find more details on how to configure the YAML in our documentation.

Next Steps

If you haven’t tried Azure Load Testing yet, you can create a resource and start testing your application today. If you are already using Azure Load Testing, you can configure the percentiles for your new tests and see the difference. Don’t miss this chance to optimize your application performance and deliver a great user experience.

Thank you for choosing Azure Load Testing!

​Microsoft Tech Community – Latest Blogs –Read More 

I have encountered the same problem as described in this thread: https://techcommunity.microsoft.com/t5/copilot-for-microsoft-365/openai-connection-with-copilot-not-returning-a-response/m-p/4025253 

After trying to change many models and GPT versions, I removed all the copilots and created a new one to check whether it would work. Unfortunately, after turning on the Generative AI option, the new copilot doesn’t even have the topic ‘Conversational Boosting’.

Has anyone else encountered this issue? If so, how did you resolve it?

PS: this copilot is created from Azure AI Studio (I want to connect the bot with the Azure OpenAI service). I tried to create another one directly on Copilot and it has ‘Conversational Boosting’ topic as normal.

​I have encountered the same problem as described in this thread: https://techcommunity.microsoft.com/t5/copilot-for-microsoft-365/openai-connection-with-copilot-not-returning-a-response/m-p/4025253 After trying to change many models and GPT versions, I removed all the copilots and created a new one to check whether it would work. Unfortunately, after turning on the Generative AI option, the new copilot doesn’t even have the topic ‘Conversational Boosting’. Has anyone else encountered this issue? If so, how did you resolve it?  PS: this copilot is created from Azure AI Studio (I want to connect the bot with the Azure OpenAI service). I tried to create another one directly on Copilot and it has ‘Conversational Boosting’ topic as normal.  Read More

​